Barclays scales Claude: 120K emails/day, 16K staff on it

Barclays is rolling Claude Code out to 50% of its developers by the end of 2026, and already uses Claude to sort roughly 120,000 client emails a day and power a knowledge assistant used by more than 16,000 colleagues. This lesson shows how to apply the same two patterns — grounded Q&A and reliable email triage — to your own work with Claude.

What shipped: Barclays, a British bank serving over 20 million UK retail customers, is expanding its use of Claude across software development and day-to-day operations. Claude Code adoption is expected to hit 50% of Barclays' developers by the end of 2026, and a majority of its software engineers by 2027.

Two things are already live, and both are patterns you can copy today.

1. A grounded knowledge assistant. Barclays' Colleague Knowledge Assistant, live since 2025, uses Claude with retrieval-augmented generation — meaning Claude answers by pulling from Barclays' own internal documents, not from memory alone. More than 16,000 colleagues use it, and it has handled over one million searches, giving staff faster answers to pass on to customers.

2. Email triage at scale. In Barclays' Global Markets business, Claude classifies, enriches, and routes incoming client emails — about 120,000 of them every day — so operations staff spend less time sorting and more time acting.

Neither use case replaces human judgment. Barclays pairs both with what Anthropic's Paul Smith called a high standard for governance: human oversight stays in the loop, and ambiguous cases still go to a person. That's the part worth copying as much as the automation itself — the win isn't "Claude decides," it's "Claude narrows the work down to what actually needs a human."

Barclays' rollout shows two reusable patterns: give Claude your own documents to answer from instead of letting it guess, and give Claude clear categories plus an explicit escalation rule when you want it to triage incoming work. Both patterns work at Barclays' scale — 120,000 emails a day, 16,000+ colleagues — and at the scale of a five-person team.